SURVEILLANCE OPTIMIZATION AS A MULTI-CRITERIA, MULTI-STAKEHOLDER PROBLEM

The primary role of a public surveillance organization is to allocate resources to surveillance activities which provide information on health status in the animal population to facilitate follow-up intervention strategies. Since the aim of hazard intervention is to maximize social welfare or minimize lack of social welfare, in this sense surveillance also indirectly contributes to social welfare optimization. The private surveillance organizations have their social responsibilities which are beyond profit maximization especially with respect to the highly socially relevant livestock hazard surveillance. Trade-offs can exist between public and private resources (e.g. when the public body monitors the surveillance activities of the private organization instead of directly conducting surveillance) and can also exist in an asymmetric distribution of these resources between stakeholders. The latter is even more prominent when the benefits of improved surveillance are concerned. For example, a reduced impact of an avian influenza (AI) epidemic because of ‘early detection of the virus’ includes mitigated human health burden, fewer animals being culled, a reduced impact on animal welfare and socio-ethics, and less disruption of social life [Reference Longworth, Mourits and Saatkamp12, Reference Longworth, Mourits and Saatkamp13]. All of these criteria are evaluated by the stakeholders involved [Reference Mourits, van Asseldonk and Huirne14]. Hence, there is a large asymmetry between stakeholders regarding both resources (i.e. costs) and benefits, which could cause conflicts of interest. The prime decision-making role of a surveillance organization is to allocate the surveillance resources to achieve collective welfare optimization given practical constraints.

HAZARD CATEGORIZATION

Surveillance organizations operate SHSSs for various hazards, all of which have specific features with regard to hazard types, surveillance objectives and occurrence possibilities. Figure 1 presents an overview of a hazard categorization.

Fig. 1.

Single hazard surveillance system categorization scheme. HRP, High-risk period.

Starting with all hazards that can, in principle, be surveyed, an initial distinction can be made between biological (viruses, bacteria, prions, etc.) and chemical (contaminants, toxins, etc.) hazards. The main reason for this distinction is the different dynamics of these hazards in an animal population. Biological hazards in principle multiply and spread between infected animals, resulting in an increasing number of affected animals over time. Most chemical hazards dilute after entering the livestock production chains (assuming only one entrance, such as a contaminated batch of feed): once entered, the concentration will reduce due to growth of the animal and/or through vertical dilution (e.g. dioxin from a sow to its offspring).

A subsequent categorization feature is prevalence. Biological hazards can be either absent in normal conditions (i.e. epidemic or zero-prevalence hazards such as CSF and foot-and-mouth disease), while prevalence cannot be excluded but is extremely low (i.e. close-to-zero prevalence hazards, such as bovine spongiform encephalopathy; BSE) or have a higher prevalence (e.g. endemic hazards such as Salmonella). For chemical hazards, higher prevalence is assumed to be non-hazardous and hence disregarded, leaving zero-prevalence (i.e. not allowed, such as added hormones) and low-prevalence (i.e. having a very low threshold, such as residuals of pesticides) hazards. However, it is noteworthy that this difference is partially artificial, caused by the current technical inability to detect.

From the prevalence situation, the ultimate surveillance objective can be derived, together with the associated technical surveillance performance indicator (TSPI) (the TSPI values are designated technical surveillance performance parameters; TSPPs). The aim of zero-prevalence hazard surveillance is to detect hazards such as CSF or foot-and-mouth disease as soon as possible from the moment of introduction in the population. This is reflected by minimizing the so-called high-risk period (HRP). Therefore, important TSPIs are the length of the HRP and the number of infected farms at the end of this HRP (as a measure for disease spread during the HRP). For close-to-zero hazards, detection of all existing cases before they pose a danger to the general public or cause significant losses for the industry (e.g. BSE) is the main objective; hence, the detection probability is an important TSPI. Higher-prevalence hazards are and will be endemic for some time. Therefore, reliable trend monitoring could be a main goal, e.g. to monitor the impact of control and reduction measures. Hence, the time lag until detection of important changes in prevalence levels and trends in this area, as well as the reliability and accuracy of this detection (i.e. sensitivity and specificity of the surveillance system), are important TSPIs. Similar objectives and TSPIs can be defined for both zero- and low-prevalence chemical hazards.

A CONCEPTUAL FRAMEWORK FOR ECONOMIC OPTIMIZATION

Surveillance organizations operate various SHSSs with limited resources. The (economically) optimal SP includes (1) those SHSSs with (2) their respective set-ups that combine to achieve maximum surveillance performance as well as the economic values of surveillance with limited resources and other constraints. Hence, a surveillance organization must make choices at two levels: (1) between SHSSs and (2) between surveillance set-ups of a SHSS. Figure 2 illustrates this decision problem. Including a SHSS in the SP automatically implies that a particular set-up, either existing or potential, must be chosen.

Fig. 2.

The conceptual framework for surveillance portfolio optimization. SHSS. Single hazard surveillance system.

This choice of a certain surveillance set-up for each SHSS results in the TSPP. Given a certain TSPP (ensured by the selected surveillance set-up), the consequential impacts of the hazard (under a given intervention strategy) can be calculated by the hazard impact simulation model or estimated by relevant experts (if the impact simulation model is missing). These impacts can be of various types (e.g. economic losses, animal welfare impacts, human health impacts, etc.) which are captured by different hazard impact indicators (HIIs). For example, for the epidemic disease AI, based on the TSPP ‘the duration of the HRP’ (under a given intervention strategy), the economic losses and annual human infections because of the AI outbreak can be estimated. The same also applies to other types of hazards. After obtaining the TSPP ‘detection probability’ for the close-to-zero disease BSE in The Netherlands, annual human deaths because of BSE infections and the level of public unease can also be calculated.

Of course, hazard impacts are not only determined by the quality of surveillance but also by the quality of intervention [Reference Howe, Häsler and Stärk15], and therefore from a comprehensive point of view, a joint analysis of surveillance and intervention is preferred [Reference Hoinville16]. However, this research's focus is on the surveillance component (i.e. we want to conduct cost-benefit analysis on surveillance only) and addressing surveillance from a portfolio perspective has already been very complex. Hence, we assume the intervention strategy is fixed (i.e. taking the default intervention strategy as given). Thus, all the differences of the avoided losses are caused by the surveillance systems, which allow us to focus on our key concern.

Moreover, surveillance costs are incurred from applying this set-up. Similarly, other hazards (i.e. SHSS) must be considered, as well as the subjective valuation by the stakeholders.

In order to elaborate the optimization problem, the following problems must be solved:

• • the various impacts each hazard has must be made comparable and additive;

• • differences in valuation of stakeholders of different impacts must be allowed, as well as interest differences between stakeholders.

Below, an attempt has been made to solve this problem in a conceptual manner in order to enable economic optimization of a SP.

Step 1. Each SP consists of a set of SHSSs, so a list of potential hazards and associated SHSSs must first be identified. Next, for each SHSS the efficient set of surveillance set-ups must be identified (details presented in [Reference Guo1]. Thereafter, for each HII i, the standardized portfolio performance should be calculated using equation (1):

(1) $${v_i} = 100 \times \displaystyle{{\mathop \sum \nolimits_{h = 1}^H \left( {P_{h,i}^0 - {P_{h,i,s}}} \right)} \over {\mathop \sum \nolimits_{h = 1}^H \left( {P_{h,i}^0 - P_{h,i}^{{\rm max}}} \right)}}\,\,\,\,\,{\rm for\,all}\,i,$$

where only one or no surveillance set-up for each hazard h is implemented;v _i denotes the standardized portfolio performance (SPP) on HII i, which is actually the performance deviation on HII i, compared to the performance of the maximum portfolio performance on HII i; P _h,i,s denotes the impact parameter (e.g. disease costs, number of human deaths) for hazard h, on HII i, for implementing surveillance set-up s; $\; P_{h,i}^0 \; $ denotes the baseline performance for hazard h, on HII i; $\; P_{h,i}^{{\rm max}} \; $ denotes the maximum performance technically possible (or artificially set by the relevant expert) for hazard h on HII i; $\; \mathop \sum \nolimits_{h = 1}^H ( {P_{h,i}^0 - P_{h,i}^{{\rm max}}} )\; $ denotes the theoretical maximum portfolio performance on HII i.

Step 2. Having obtained the overall performance of the entire SP for each HII i, two subjective weightings must be performed: (1) the differences in preference between the stakeholders involved, and (2) the differences in importance of the various stakeholders viewed by the final decision maker. This ‘double weighting’, as well as the final optimization statement, is expressed in equations (2)–(5):

(2) $$\eqalign{{\rm Max\;} PV{\rm \;} & = \left( {\matrix{ {{w_1}} & \cdots & {{w_G}} \cr}} \right) \cr & \quad \times \left( {\matrix{ {{w_{1,1}}} & \cdots & {{w_{1,I}}} \cr \vdots & \ddots & \vdots \cr {{w_{G,1}}} & \cdots & {{w_{G,I}}} \cr}} \right) \cr & \quad \times \left( {\matrix{ {{v_1}\left( X \right)} \cr \vdots \cr {{v_I}\left( X \right)} \cr}} \right),}$$

s.t. various constraints, such as

(3) $$\sum\nolimits_{s = 1}^{{S_h}} {{x_{h,s}} \le 1} \,\,\,\,{\rm for\,all}\,h,$$

(4) $$\sum\nolimits_{h = 1}^H {\sum\nolimits_{s = 1}^{{S_h}} {{c_{g,h,s}}{x_{h,s}} \le {B_g}}} \,\,\,\,{\rm for\,all}\,g,$$

(5) $${x_{h,s}} \in (0,\; 1)\,\,\,\,{\rm for\,all}\,h,s,$$

where PV is the overall weighted portfolio performance (OWPP); (w ₁ ··· w _G ) is the weights the decision maker places on the various stakeholders 1 to G; $\left( {\matrix{ {{w_{1,1}}} & \cdots & {{w_{1,I}}} \cr \vdots & \ddots & \vdots \cr {{w_{G,1}}} & \cdots & {{w_{G,I}}} \cr}} \right)$ are the preference weights stakeholders 1 to G place on HII 1 to I.

$\left( {\matrix{ {{v_1}\left( X \right)} \cr \vdots \cr {{v_I}\left( X \right)} \cr}} \right)\; $ is the SPP on each HII from equation (1);

X is the decision variable matrix of x _h,s . x _h,s denotes the binary decision variable to judge whether, for hazard h, surveillance set-up s is selected to compose the SP. S _h denotes the number of the alternative surveillance set-ups for hazard h. B _g denotes the total annual budget available for stakeholder group, g (g = 0, 1, …, G), to carry out the surveillance activities (to simplify the formulation, the decision maker (the surveillance organization, e.g. FSA) is treated as a stakeholder group, g = 0). c _g,h,s is the annual surveillance costs for stakeholder group g, when, for hazard h, surveillance set-up s is implemented.

The set of constraints [equation (3)] ensures that a maximum of one surveillance set-up for each hazard will be included in the SP; constraints [equation (4)] ensure that the total annual surveillance costs for stakeholder group g cannot exceed the annual available surveillance budget available; and definitions [equation (5)] define x _h,s as binary variables. Additional constraints can be included to establish more complex models according to the specific situation; such as the minimum required surveillance performance for some hazards.

To derive the weight matrix that reflects shareholders’ preferences, a stakeholder panel, analogous to the consumer panel in marketing (e.g. [Reference Dynan17–Reference Booth19]) is established to elicit stakeholder preference.

Quantitative elaboration of the concept of SP optimization

To illustrate the concept of optimization of a SP, we use a hypothetical numerical example. Three different hazards were selected as potential surveillance targets within the portfolio: CSF, AI, and Salmonella. Impact parameters for Dutch conditions were assumed referring to previous studies (AI [Reference Backer20, Reference Koopmans21], and Salmonella [Reference Valkenburgh22]). These impact parameters, which mimic the results of SHSS analysis, are presented in Table 1.

Table 1.

Impact parameters, P_his, for hazard h, on indicator i, for implementing surveillance setup s, as well as weights for inidcators and stakeholders

FSA, Food safety authority.

Impact parameters are fictively generated referring to the works of Backer et al. [Reference Backer20], Koopmans et al. [Reference Koopmans21], Mangen et al. [Reference Mangen23], and Valkenburgh et al. [Reference Valkenburgh22] under the non-vaccination and pre-slaughter control strategy.

Four surveillance set-ups are included for each hazard. For each, respective surveillance costs for an FSA and the impact on disease costs and human health are listed according to the framework in Figure 2. Table 1 lists the weights for each HII and for farmers and citizens. The impact parameters on each HII are fictively generated, given the aim of this research is simply to demonstrate the rationale of the framework. Some of the data refer to [Reference Backer20–Reference Valkenburgh22] which are assumed to reflect some impacts of the default surveillance situation. In this way, we try to make the whole data fall in a relatively reasonable range. For example, the costs due to a CSF outbreak in the default situation (i.e. surveillance set-up 1) are according to [Reference Mangen23] (see also the appendix of [Reference Guo1]). The information we use from [Reference Backer20, Reference Koopmans21] is that there is one human death caused by AI since 2003, and therefore the annual death rate is approximately 0·1. The information we use from [Reference Valkenburgh22] is that in 2004 the number of human salmonellosis cases was 35000, and an estimate that 39 people eventually died. The cost-of-illness because of human salmonellosis was estimated as €7 million per year.

Step 1. Table 2 lists the SPPs, v _i , of the possible SPs. It is assumed here that one surveillance set-up must be selected for each hazard, which means there are a total of 64 SPs available. For ease of demonstration, only four of these SPs are explicitly shown in Table 2.

Table 2.

Standardized portfolio performance per indicator

FSA, Food safety authority.

Step 2. Through the ‘double weighting’ on v _i , the OWPP for all 64 possible SPs are obtained and graphically depicted in Figure 3.

Fig. 3.

Overall weighted portfolio performances (OWPPs) with three different sets of decision makers’ weights on two stakeholder groups. FSA, Food safety authority.

The horizontal axis represents the annual surveillance costs for Dutch FSAs to operate the SP and the vertical axis represents the OWPP. Each SP is defined by (1) the annual surveillance costs and (2) the OWPP. Clearly, there is no proportional relationship between the OWPPs and the costs, which articulates the need for economic optimization of surveillance resource allocation. Figure 3(a–c) present the results with three different settings of decision makers’ weights on farmers and citizens. Two levels of budget constraints (X1 = €36 million or X2 = €100 million) are considered and only the SPs that expend less than the budget are feasible options. Similarly, two minimum performance constraints (Y1 = 60% or Y2 = 90%) are also considered to ensure an acceptable level of OWPP; in other words, only those SPs that guarantee the minimum OWPP are feasible.

Figure 3a shows the results solely from a citizen's point of view, where the decision makers’ weights on farmers and citizens are 0 and 1 under the budget constraint. In terms of cost-effectiveness, and taking into account only the preference of the citizens, SP-a is to be preferred if case budget constraint X1 is considered. An increase of this budget to X2 results in a switch in preference to SP-b. A subsequent step could be to treat the cost-effective SPs (i.e. SP-a and SP-b) as a starting point for cost-efficient analysis. For budget constraint X1, from the cost-efficient perspective, SP-d can also be an attractive option because it delivers slightly lower overall OWPPs than SP-a but saves on surveillance costs. For budget constraint X2, SP-c is even more attractive than SP-b because it saves almost one-third on surveillance costs but still ensures almost the same OWPPs as SP-b. Such an analysis has practical implications for decision makers to efficiently allocate surveillance budgets. Another approach could be to take the minimum performance as a constraint. For example, if the minimum required OWPP is Y1, SP-d is the cheapest choice to fulfil the requirement, while if the minimum required OWPP increases to Y2, the least expensive SP to fulfil that requirement becomes SP-c.

Conducting the same analysis purely from a farmer's standpoint (Fig. 3b ) produces different results. Considering budget constraint X1, SP-e is preferred over SP-a. Moreover, the OWPP under SP-a becomes smaller, indicating that farmers have a lower overall preference for this surveillance set-up. Similarly, with constraint Y1, SP-f rather than SP-d is the cheapest SP for ensuring the minimum performance, while with Y2, SP-g rather than SP-c is the cheapest SP. Only with budget constraint X2 do both farmers and citizens prefer SP-b.

Figure 3c shows the results from a simultaneous viewpoint of both stakeholder groups based on the level of importance judged by the decision makers. The preferred SP-a and SP-b under budget constraints X1 and X2 are the same as in Figure 3a because the decision maker assigns a larger weight on citizens, which means that their preference is relatively dominant. Moreover, the cheapest SP under minimum performance constraint Y1 is SP-a, which is different from its counterparts in Figure 3a (SP-d) and Figure 3b (SP-g). As in Figure 3a , SP-c is the cheapest option for satisfying constraint Y2. This compromise-based result provides the scientific basis for decision makers to arrive at their decisions.

For surveillance organizations such as a FSA, it should be realized that the complexity of the SP optimization problem increases with (1) the number of hazards, (2) the level of surveillance set-ups for each hazard, and (3) the number of stakeholder groups involved.